Texture-aware and defect-guided swin transformer for multi-scale textile defect segmentation

Fabric surface defect segmentation is a crucial yet challenging task in industrial visual inspection, where complex textures, low contrast, and diverse defect patterns often degrade detection accuracy. To address these challenges, this study introduces a novel Swin Transformer-based segmentation framework that integrates texture enhancement, defect-aware representation learning, and hybrid multi-scale context modelling. Specifically, a Texture Enhancement Module (TEM) is introduced to strengthen low-level texture representations by capturing global statistical distributions, alleviating the ambiguity caused by weak contrast in fabric textures. Furthermore, a Defect-Guided Attention (DGA) mechanism is designed to emphasise defect-relevant regions while suppressing background interference, leading to more discriminative feature learning. To comprehensively exploit multi-scale information, we propose a Hybrid Multi-scale Context (HMC) module that combines complementary context extraction strategies to capture fine-grained spatial details and global semantic information from hierarchical features. The enhanced features are finally decoded using a UPerNet-style decoder for precise pixel-level prediction. Extensive experiments conducted on the ZJU-Leaper textile defect dataset demonstrate that the proposed method consistently outperforms state-of-the-art segmentation approaches across four representative fabric groups especially under complex texture conditions, verifying its robustness and effectiveness for industrial fabric inspection.